Oil spills are considered one of the major threats to the marine and coastal environment. Synthetic aperture radar (SAR) sensors are frequently employed for this purpose due to their ability to operate effectively under various weather and illumination conditions. SAR can clearly capture oil spills with distinctive radar backscatter intensity, resulting in dark regions in the images. This characteristic enables the monitoring and automatic detection of oil spills in SAR imagery. U-Net stands as one of the commonly employed semantic segmentation models, known for its ability to achieve superior segmentation performance even with limited training data. In this study, a modified lightweight U-Net model was introduced to enhance the performance of maritime multi-class segmentation in SAR images. First, a lightweight MobileNetv3 model served as the backbone for the U-Net encoder to perform feature extraction. Secondly, the convolutional block attention module (CBAM) was employed to enhance the network's capability in extracting multiscale features and to expedite the module calculation speed. The experimental results showed that the detection accuracy of the proposed method can achieve 77.07% of the mean Intersection-Over-Union (mIOU). Compared with the original U-Net model, the proposed architecture can improve the mIOU about 4.88%.

With the characteristics of high resolution, strong penetration ability, all-day observation and wide spatial coverage, Synthetic Aperture Radar (SAR) image has been widely used in shoreline detection. However, the shoreline detection will be affected by the shadows of geometric distortion caused by the side-looking of SAR especially in areas with large terrain fluctuations, such as the eastern coast of Taiwan. Therefore, this study proposed an efficient shoreline detection method using dual-polarization Sentinel-1 SAR and digital elevation model (DEM) data based on deep learning methods. In this research, two common self-built datasets were introduced, which covered all coastal areas of the study area in Taiwan island. The datasets included a total of 4,029 and 3,522 images, respectively. One contains VH polarization images and the other consists of a three-layer stack with VH, VV polarization images and DEM data. The training images of the dataset were labeled by manual inspection and morphological processing. In this study, the shoreline detection was based on the semantic segmentation U-Net model with batch normalization (BN) module. The segmentation results of the U-Net model were then processed by morphological postprocessing and edge detection for shoreline detection. Experimental results show that the combination of dual-polarization SAR and DEM data significantly improves shoreline detection results compared to those using Sentinel-1 VH imagery.

Natural disasters are causing unpredictable and devastating effects on people and their property worldwide as extreme weather events become more frequent and severe due to global climate change. Early assessment of disaster situations by relief teams can help to minimize the loss of life and property. Remote sensing satellite imagery has been used to estimate disaster losses and reduce the time and cost involved in damage assessments. However, the accuracy and performance of deep learning methods used to assess damage to buildings still leave room for improvement. In this research, we propose a novel deep learning model architecture, ASPP-Attention-ResNeSt-Unet (AARNS-Unet), to identify building areas and assess damage using remote sensing satellite images taken before and after a disaster. We evaluate the performance of our model using the public dataset provided by the X-View2 competition. Our proposed model combines ResNeSt, Atrous Spatial Pyramid Pooling (ASPP), and Convolutional Block Attention Module (CBAM) for better generalization. Our experiments show that the proposed model achieves a 1.3% accuracy improvement and a 7-time reduction in training time compared to other deep learning methods. Our proposed method can aid disaster relief teams in evaluating the damage to buildings and informing their response efforts. This research contributes to the development of more accurate and efficient methods for disaster assessment, which can help to reduce the negative impact of natural disasters on society.

This study proposed a paddy rice mapping based on hand-crafted features combined with deep learning methods. In this research, the rice growth-related features were extracted from time-series SAR data provided by C-band Sentinel-1A images. Four rice features were first extracted from the rice growth curve, including Average Normalized Backscatter (ANB), Backscatter Variation Rate (BVR), Time Interval (TI) and Backscatter Difference (BD). Then, the deep learning U-Net model with four combined rice features was used to obtain the mapping distribution of paddy rice-fields. In the study, the experimental areas were composed of two important rice growth counties in central Taiwan, including Yunlin and Changhua counties. The experimental results showed that the detection accuracy of the proposed method for rice and non-rice can achieve 92.3% and 98.2%, respectively. These results demonstrate the great potential of SAR data in mapping paddy fields using the U-Net model with proposed feature inputs.

Taiwan's agriculture is an important national economic industry. Ensuring food security and stabilizing the food supply are the government's primary goals. The Agriculture and Food Agency (AFA) of the Executive Yuan's Council of Agriculture has conducted agricultural and food surveys to address those issues. Synthetic aperture radar (SAR) images will not be affected by climatic factors, which makes them more suitable for the forecast of rice production. This research uses the spatial-temporal neural network convolutional long short-term memory network (ConvLSTM) to identify rice fields from SAR images. The results show that ConvLSTM can greatly reduce the proportion of model false positives to 51.16%, produced higher average precision of 95.70%, and F1-score of 0.9648. The ConvLSTM neural network has produced good results for rice field identification compared with state-of-the-art neural networks.

Synthetic Aperture Radar (SAR) imagery has been widely used in many maritime applications due to its high resolution, wide coverage, and real-time monitoring characteristics. Nevertheless, the size of SAR images is significantly large for real-time application. In recent years, High-Performance Computing (HPC)-related methods have been used to improve the precision and detection rate of SAR imagery analysis. In this paper, motivated by the state-of-the-art real time object detection You Only Look Once version 3 (YOLOv3), an enhanced GPU-based deep learning method has been proposed, namely Accelereated-YOLOv3 (A-YOLOv3), to detect ships from the SAR images. A-YOLOv3 aims to reduce the computational time with relatively competitive detection accuracy by constructing a new architecture with less layers and channels. The proposed A-YOLOv3 architecture achieves Average Precision (AP) of 97.4% on the Expand Diversified SAR Ship Detection Dataset (EDSSDD).

Ship detection is one of the most important researches in the field of navigation safety and marine environment monitoring. Synthetic aperture radar (SAR) imagery has been used as a promising data source for monitoring maritime activities. However, the resolution of SAR images is limited, and it cannot effectively detect densely distributed and small ships, especially near harbors. In order to effectively manage ships during the day and night, this research uses visible and infrared images for ship detection. In this study, an improved architecture based on You only look once version 3 (Yolov3) is proposed for ship detection. Yolov3 provided multi-scale feature extraction to enhance the recognition of small targets. In addition, the study also considered the influence of Yolov3 parameters on ship detection, including the input image size, the number of filters in convolution layers and the detection scales. The experiment is based on a data set containing six types of ships, and a total of 5, 513 visible and infrared images from the harbors in northern Taiwan. The experimental results show that when the model parameters are selected as: 352x352 input size, the scale of large target deleted and the convolution filters reduced by 30%, Yolov3 has better ship detection performance and computational efficiency. Compared with the original Yolov3 with 87.9% mean average precision (mAP) and 87.0 billion floating point operations per second (BFLOPs), the proposed architecture can achieve 89.1 % mAP and 24.3 BFLOPs.

Rice is the most important food source for many countries, which especially for Asia, includes Taiwan. Monitoring rice field distribution can effectively manage food security. This study proposed a feature-based decision approach to detect the mapping of rice cultivation using the time-series Synthetic Aperture Radar (SAR) data provided by Sentinel-1A. Instead of using the maximum and minimum backscatter of SAR data, as most studies in the literature, this study established a rice growth model based on complete time series data in the rice growth period. From the developed model, a feature related to rice growth time, that is, the time interval (TI) between vegetative growth and maturity stages was introduced. The proposed feature was first compared with the feature of backscatter difference (BD) between the maximum and minimum value of SAR data. The experimental results show that the proposed feature can achieve better rice detection accuracy. Then, a decision method based on the combination of TI and BD features was proposed for rice planting mapping. In the study, Yunlin and Changhua counties in central Taiwan were used as experimental areas. The experimental results show that the proposed method can achieve more than 90% overall accuracy in rice detection for VH polarization. Furthermore, comparing with the traditional method that uses growth height feature, BD, the proposed method can improve the overall accuracy of rice detection about 5%.

Rice is one of the most important and valuable crops in the world. People around the world mainly depend on rice as their daily diet. Therefore, efficient rice fields monitoring is a crucial factor in the improvement of rice crop yield estimation, damage evaluation, budget planning, and agricultural resource management. Synthetic aperture radar (SAR) is an effective tool in monitoring agricultural fields because of its ability to provide high resolution images regardless of weather conditions. However, precision agriculture has put higher requirements for SAR data analysis. In recent years, deep learning methods have achieved great success in various remote sensing applications. In this research, two of the most popular deep learning architectures for time series data, namely convolutional long short-term memory (ConvLSTM) and gated recurrent unit (GRU) have been explored and applied to detect rice fields from SAR images in Taiwan. The experimental results showed that time-series deep learning methods for analyzing SAR data have a great potential for improving the rice fields detection.

In recent years, the satellite technique has a tremendous progress. The images captured by satellites contain larger data and dimensions. The higher number of spectral bands increases the complexity of a classification task. Therefore, it is necessary to reduce highly correlated and redundant neighboring bands which cause the huge phenomenon. In this paper, we proposed a hybrid hierarchical approaches that combine the greedy modular eigenspace (GME) and impurity function band prioritization with hotspot analysis. Unfortunately, GME doesn't guarantee to reach a global optimal solution by the greedy algorithm except by the exhaustive search method. In order to mitigate this limitation, we used a particle swarm optimization (PSO) algorithm to cluster the highly correlated bands and hotspot analysis to give weighting to the clustered blocks. The experimental results on two publicly available benchmark dataset demonstrate that the presented approach can select those bands with discriminative information. The effectiveness of the proposed approach is tested on both images with different parameters of PSO. To verify the effectiveness of a hybrid hierarchical approach put forward in this paper, KNN classifier is performed on the selected bands. In MASTER dataset, the proposed method has 90.91% achievement in dimensionality redaction rate with classification accuracy of 95.3%. In Northwest Tippecanoe County (NTC) dataset, the dimensionality reduction rate is 87.7% and the method achieve a classification accuracy of 96.48%.The results clearly show that the proposed method has the better effects both in dimensionality reduction rate and classification accuracy.

In this paper, a novel technique, known as a modified adaptable nearest feature space (MANFS) classifier, is proposed for supervised classification of remote sensing images. The original nearest feature space (NFS) may cause misclassification if the test samples are close to the different class training samples which are highly overlapped. Thus it is difficult to discriminate different classes. Compared to the original NFS classifier, we propose a novel MANFS classifier, which can precisely analysis the coverage of the feature space used in NFS and perfectly confine the extensible ranges of each feature space of NFS, to reduce the impact of the overlapping training samples of different categories. Experimental results show that MANFS achieves better classification accuracy than the original NFS one for remote sensing image.

Modern satellite imaging technology has resulted in an increased number of hyperspectral bands acquired by state-of-the-art sensors. It significantly advances the field of remote sensing. Owing to the increasing number of bands, the huge data quantity causes the curse of dimensionality and leads to the worse accuracy. It also increases the computational complexity exponentially as the problem size increases. It's therefore important to reduce dimensionality in order to prevent the curse of dimensionality. In this paper, a novel dimensionality reduction, named impurity function band prioritization method based on the particle swarm optimization and the gravitational search algorithms, is proposed to reduce the number of hyperspectral bands. The experimental results show that our approach can efficiently reduce dimensionality of hyperspectral data sets and significantly achieve a better classification accuracy compared to other methods.

This paper presents a promising band selection algorithm, known as particle swarm optimization/impurity function class overlapping (PSO/IFCO) method, which adopts a novel multiple attribute decision making (MADM) model approach to the hyperspectral remote sensing images. The proposed MADM-based PSO/IFCO method can be divided into two steps: 1) PSO algorithm and 2) the IFCO scheme. With PSO band selection algorithm, the highly correlated bands of hyperspectral imagery can first be grouped into band modules, known as greedy modular eigenspace (GME), to coarsely reduce high-dimensional datasets in the first step. The more highly correlated small modules are further constructed with the statistics of impurity weights calculated by IFCO scheme in the second step. These statistics results of impurity weights are used to finely select the most important feature bands from the hyperspectral imagery. The proposed MADM-based PSO/IFCO makes use of the correlation coefficients matrix to cluster the highly correlated bands together and obtain GME in the first step. More specifically, we use the analytic hierarchy process (AHP) model, which is the most suitable implementation of MCDM for proposed method, to examine hierarchically the relations among different GME modules with the impurity weighted by IFCO in the second step. Finally, by accommodating the statistics of impurity weights, the proposed MADM-based PSO/IFCO method can effectively select the most representative features for hyperspectral band selection and reduction. The effectiveness of the proposed method is evaluated by MASTER and AVIRIS hyperspectral images. The experimental results demonstrate that the proposed method can not only enhance the high dimension reduction rate, but also offer a satisfactory classification performance.

In this paper a novel technique based on nearest feature space (NFS), known as incenter-based nearest feature space (INFS), is proposed for supervised hyperspectral image classification. Due to the class separability and neighborhood structure, the traditional NFS can perform well for classification of remote sensing images. However, in some instances, the overlapping training samples might cause classification errors in spite of the high classification accuracy of NFS for normal cases. In response, the INFS is proposed to overcome this problem in this paper. INFS method makes use of the incircle of a triangle which is tangent to its three sides and form a INFS. In addition, an incenter can be calculated by three training samples of the same class efficiently. Furthermore, in order to speed up the computation performance, this paper proposes a parallel computing version of INFS, namely parallel INFS (PINFS). It uses a modern graphics processing unit (GPU) architecture with NVIDIA's compute unified device architecture (CUDA) technology to improve the computational speed of INFS. Experimental results demonstrate the proposed INFS approach is suitable for land cover classification in earth remote sensing. It can achieve the better performance than NFS classifier when the class sample distribution overlaps. Through the computation of GPU by CUDA, we can also gain better speedup.

In the study, an efficient compression approach, group and region based KLT (GR-KLT), is proposed for hyperspectral imagery. The GR-KLT contains one clustering signal subspace projection (CSSP) segmentation method and the maximum correlation band clustering (MCBC) method. The CSSP first divides the image into proper regions and the MCBC partitions the spectral bands into several groups according to their associated band correlation for each image region. By the way, the image is further compressed by the KLT-JPEG for each group in each image region. Furthermore, we develop a parallel architecture for the GR-KLT compression algorithm. Simulation results performed on AVIRIS images have demonstrated the efficiency of the proposed approaches.

In the study, we propose a novel unsupervised classification technique for hyperspectral images, which consists of two algorithms, referred to as the maximum correlation band clustering (MCBC) and hierarchical binary quaternion-moment-preserving (BQMP) thresholding technique. By the MCBC, we partition the bands into groups and transfer the high-dimensional image data into low-dimensional image features. Afterwards, the hierarchical BQMP approach partitions the feature image into proper regions according to the spectral characteristics. Simulation results performed on AVIRIS images have demonstrated the efficiency of the proposed approaches.

Greedy modular eigenspaces (GME) has been developed for the band selection of hyperspectral images (HSI). GME attempts to greedily select uncorrelated feature sets from HSI. Unfortunately, GME is hard to find the optimal set by greedy operations except by exhaustive iterations. The long execution time has been the major drawback in practice. Accordingly, finding an optimal (or near-optimal) solution is very expensive. In this study we present a novel parallel mechanism, referred to as parallel particle swarm optimization (PPSO) band selection, to overcome this disadvantage. It makes use of a new particle swarm optimization scheme, a well-known method to solve the optimization problems, to develop an effective parallel feature extraction for HSI. The proposed PPSO improves the computational speed by using parallel computing techniques which include the compute unified device architecture (CUDA) of graphics processor unit (GPU), the message passing interface (MPI) and the open multi-processing (OpenMP) applications. These parallel implementations can fully utilize the significant parallelism of proposed PPSO to create a set of near-optimal GME modules on each parallel node. The experimental results demonstrated that PPSO can significantly improve the computational loads and provide a more reliable quality of solution compared to GME. The effectiveness of the proposed PPSO is evaluated by MODIS/ASTER airborne simulator (MASTER) HSI for band selection during the Pacrim II campaign.

In this paper we present a parallel classification learning method, referred to as parallel minimum classification error (PMCE) learning, for supervised classification of multisource remote sensing images. The approach is based on the positive Boolean function (PBF) classifier scheme. The PBF implements the minimum classification error (MCE) as a criterion to improve classification performance. By evenly distributing both positive and negative samples of MCE learning modules to different PMCE learning nodes, PMCE outperforms the original one in terms of execution time. It fully utilizes the significant parallelism embedded in MCE learning of PBF to create a set of PMCE learning nodes implemented by using the message passing interface (MPI) library and the open multi-processing (OpenMP) application programming interface. A sophisticated hierarchical structure of hybrid PMCE, which combines cluster based MPI with multicore-based OpenMP, is proposed to demonstrate the flexibility of implementation of the proposed scheme. The effectiveness of the proposed PMCE is evaluated by fusing MODIS/ASTER airborne simulator (MASTER) hyperspectral images and the Airborne Synthetic Aperture Radar (AIRSAR) images for land cover classification during the Pacrim II campaign. The experimental results demonstrated that PMCE can improve the computational speed of PBF classification significantly.

In this paper we present a parallel band selection approach, referred to asparallelsimulatedannealingbandselection(PSABS), for hyperspectral imagery. The approach is based on thesimulatedannealingbandselection(SABS) scheme. The SABS algorithm is originally designed to group highly correlated hyperspectral bands into a smaller subset of band modules regardless of the original order in terms of wavelengths. SABS selects sets of non-correlated hyperspectral bands based onsimulatedannealing(SA) algorithm and utilizes the inherent separability of different classes in hyperspectral images to reduce dimensionality. In order to be effective, the proposed PSABS is introduced to improve the computational speed by using parallel computing techniques. It allowsmultipleMarkovchains(MMC) to be traced simultaneously and fully utilizes the significant parallelism embedded in SABS to create a set of PSABS modules on each parallel node implemented by themessagepassinginterface(MPI) cluster-based library and theopenmulti-processing(OpenMP) multicore-based application programming interface. The effectiveness of the proposed PSABS is evaluated byMODIS/ASTERairbornesimulator(MASTER) hyperspectral images for hyperspectral band selection during the PACRIM II campaign. The experimental results demonstrated that PSABS can significantly improve the computational loads and provide a more reliable quality of solution compared to the original SABS method.

In the study, we proposed two adaptive classifiers based on image eigen-features for multispectral image classification. An adaptive signal subspace projection (ASSP) approach is first proposed to detect and extract target signatures in unknown background. The weights of ASSP are adjusted adaptively by using the eigen-features which are updated recursively by the adaptive eigen-decomposition algorithm. Then, we proposed an artificial neural networks (ANN) based on back propagation multilayer perception (BPMLP) with weights trained by the image eigen-features. Simulation results validate the image eigen-features can alleviate the noise effect in classification and the proposed ASSP and BPMLP classifiers have lower detection error and fast convergence rate than conventional Wiener filter and per-pixel ANN methods.

In this study, we propose an efficient approach, which consists of an algorithm to speed up the multi-dimensional image segmentation and an efficient automatic mechanism to manage satellite image databases. Concerning image segmentation, we develop a 1D segmentation technique by an eigen-subspace projection approach, which can perform the image segmentation efficiently by transforming the multi-dimensional image data into 1D projection length. In addition, we develop a data model of the eigen-index by a probabilistic approach. Based on the maximum a posterior probability (MAP) information criterion, the management scheme may determine the accuracy of the user added data entry. Therefore, the DB archive can be managed efficiently. Simulation results performed on SPOT images have demonstrated the proposed approach is suitable for management of satellite image databases

In the study, we propose an efficient projection-based segmentation technique for multi-spectral image. By projecting the multi-spectral image onto a referenced subspace, we transform the multi-dimensional (MD) image data into one-dimensional (1D) projection length. One proper referenced subspace with the largest deviation from the mean image signature vector is suggested in the study. After the transformation, any efficient 1D segmentation technique, such as moment-preserving method, can be applied. In the segmentation, we perform the projection procedure recursively and partition the image into proper regions according to their spectral characteristics. Then, to get a more integrated segmentation result, we adopt N nearest neighbor rule to merge the image regions according to their spatial correlation. Simulation results performed on SPOT and Landsat images have demonstrated the efficiency of the proposed approach. In addition, the segmentation result is suitable for indexing of satellite image databases

In the study, we present a novel indexing technique for the retrieval of satellite image databases. The approach partitions the original image into some eigen-regions according to local terrain characteristics. And index for each eigen-region, called eigen-index, is extracted by the principal eigenvector of region. Using the eigen-index separation between the specified cover type of query image and that of each object image in the database, the access of archive could be facilitated. To show the performance of the proposed approach, a PHP-based interface agent using the eigen-index for content-based access to a test archive of SPOT images is available over the internet [http://dsp.mmd.ntou.edu.tw]. Through this agent, a user can browse ten best-matched candidates for each query.

A natural, and readily computable, first guess at a solution to the coin changing problem is the canonical solution. This solution is a special case of the greedy solution which is a reasonable heuristic guess for the knapsack problem. In this paper, efficient tests are given to determine whether all greedy solutions are optimal with respect to a given set of knapsack objects or coin types. These results improve or extend previous tests given in the literature. Both the incomplete and complete cases are considered.